The MOOC Pivot
Justin Reich and José A. Ruipérez-Valiente (MIT), Science 363(6423):130–131, 2019. Read here from the author post-print in MIT DSpace, extracted locally. T1 — peer-reviewed, and the first source in this spoke that reports what happened to learners rather than how many people starred a repository.
The dataset is the argument: every MITx and HarvardX MOOC on edX from the start of the initiative to May 2018 — 565 course iterations across 261 distinct courses, 12.67 million registrations from 5.63 million learners. Harvard and MIT publish this data; most online-education operators do not, which is why the paper exists in this form at all.
Three findings
Almost nobody comes back. The largest cohort was year 4 (2015–2016) at 1.1 million unique learners, and 12% of them took another HarvardX or MITx course the following year. Second-year retention fell with every single cohort: 38% in the first cohort to 7% by the 2016–2017 cohort. Cohorts after 2015–2016 are half the size of the peak.
Growth went to the already-rich. Splitting registrants and certificate-earners by the UN Human Development Index quartile of their home country, nearly all of the growth — and the later decline — is in the most affluent quartile. The authors' verdict: MOOCs are "primarily a complementary asset for learners within existing systems," not a new pathway at the margins.
Completion did not improve. Reported for three non-exclusive groups — all participants, learners who state on a survey that they intend to complete, and learners who pay for a verified track. "Six years of investment in course development and learning research has not produced meaningful improvements in these figures." Note the middle group: filtering down to people who declared an intention to finish did not rescue the number.
Provenance caveat: the per-year completion percentages live in Figure 3, an image the local PDF extraction cannot read. The flat-over-six-years finding and the three-group structure are quoted from the body text; the specific percentages are not reproduced here because this wiki has not seen them.
The conclusion is a business-model one. edX, Coursera and FutureLearn have pivoted to helping universities outsource online professional master's degrees — the "school-as-a-service" model that Pearson Embanet, 2U and Wiley have run for two decades — competing mainly by cutting labour costs through automation.
What it does to this spoke's central hole
synthesis has carried "Does anyone finish?" as the corpus's central hole since the spoke was founded, with the standing caveat that a star count measures intent to learn, not learning. This is the first evidence in the spoke that bears on it, and it is not free.
What it does not do: a MOOC is not a meta-curriculum. coding-interview-university, developer-roadmap and awesome-system-design-resources are repositories with no enrolment, no cohort, no certificate and no completion event to measure. Nothing here transfers by arithmetic, and a reader who quotes "7% retention" about a GitHub curriculum is misusing this paper.
What it does do: it removes an excuse. The nearest measurable population to self-directed online learners — same subjects, same delivery, overlapping audience, and better instrumented, since a MOOC at least knows who registered — shows collapsing return rates and completion that six years of deliberate effort could not move. The optimistic reading of a 352k-star repository requires believing that an artifact with less structure, no deadlines and no instructor produces better follow-through than one with all three. That reading is now the one carrying the burden of proof.
It also speaks to "Who are these for?" The MOOC answer, measured, is: already-educated people in affluent countries, using it as a complement to an education they already have. That is a live hypothesis for the star counts too, and the spoke should stop treating "enormous population of self-taught entrants" as the default reading.
The supply side, eight years on (added 2026-09-09)
harvard-free-courses-catalog is what one half of this paper's dataset looks like today: 129 free courses, all online, all still on edX, with verified certificates at $219–299. The form did not retreat after the finding — it is being published at the same scale, by the same institution, under the same business model.
Two things the comparison makes visible. The catalogue's subject mix is a faculty's shape, a third humanities and three mathematics courses, which is not what a demand-driven offer looks like and fits the paper's "complementary asset for learners within existing systems" reading. And Harvard shows a prospective student none of this paper's numbers — no enrolment, no completion, no retention anywhere in the catalogue. The institution that instrumented itself well enough to publish the collapse does not put it in front of the people deciding whether to enrol.
Related
harvard-free-courses-catalog · mooc · meta-curriculum · coding-interview-university · developer-roadmap · awesome-system-design-resources · self-directed-disposition-study · justin-reich · interview-preparation · synthesis